CN116863277A

CN116863277A - RPA-combined multimedia data detection method and system

Info

Publication number: CN116863277A
Application number: CN202310933160.0A
Authority: CN
Inventors: 吕国伟; 李超; 王丽; 赵银龙; 宋占元
Original assignee: Beijing Zhongguancun Kejin Technology Co Ltd
Current assignee: Beijing Zhongguancun Kejin Technology Co Ltd
Priority date: 2023-07-27
Filing date: 2023-07-27
Publication date: 2023-10-10

Abstract

The embodiment of the application provides a method and a system for detecting multimedia data by combining RPA (remote procedure A). When a first number of multimedia interactive image sequences carrying priori abnormal violation feature data and a second number of multimedia interactive image sequences not carrying priori abnormal violation feature data are obtained, unsupervised network learning optimization is firstly carried out on a basic abnormal violation detection network, a third number of times of the first abnormal violation detection network after optimization is recycled, a final second abnormal violation detection network is generated, then first fuzzy abnormal violation feature data are generated, and the second abnormal violation detection network is trained by combining the multimedia interactive image sequences carrying priori abnormal violation feature data with the multimedia interactive image sequences not carrying priori abnormal violation feature data and the corresponding first fuzzy abnormal violation feature data, so that the training speed of the abnormal violation detection network is improved, and the workload of training annotation is reduced.

Description

RPA-combined multimedia data detection method and system

Technical Field

The embodiment of the application relates to the technical field of artificial intelligence, in particular to a method and a system for detecting multimedia data by combining RPA.

Background

Robot process automation (Robotic Process Automation, RPA) is a business process automation technology based on software robots and artificial intelligence. The RPA system is an application program that provides another way to automate the end user manual process by mimicking the end user manual process at a computer. With the development of internet multimedia data platforms, various multimedia data platforms based on mobile internet technology have been developed, however, in the interaction flow of multimedia data, some users may have some abnormal violations, so that it is required to detect the abnormal violations of the users in time and perform corresponding processing, such as shielding, number sealing, and the like. In the related art, training and scheduling of an artificial intelligent model can be generally performed by combining an RPA (remote procedure analysis) system, so that abnormal violation data existing in a multimedia interactive image are detected, however, in the related art, when training of an abnormal violation detection network is performed, a large number of priori abnormal violation feature data are required to be marked, the workload of training marking is large, and the training speed of the abnormal violation detection network is slow.

Disclosure of Invention

In order to at least overcome the above-mentioned shortcomings in the prior art, an object of an embodiment of the present application is to provide a method and a system for detecting multimedia data in combination with RPA.

According to an aspect of the embodiment of the present application, there is provided a method for detecting multimedia data in combination with RPA, including:

acquiring a first number of multimedia interactive image sequences carrying priori abnormal violation feature data and a second number of multimedia interactive image sequences not carrying priori abnormal violation feature data;

performing unsupervised network learning optimization on the basic abnormal violation detection network according to the multimedia interactive image sequence which does not carry priori abnormal violation feature data, and generating a first abnormal violation detection network;

performing network learning optimization for the first abnormal violation detection network traversal for a third number of times according to the multimedia interactive image sequence carrying the priori abnormal violation feature data, and generating a second abnormal violation detection network;

generating first fuzzy anomaly violation feature data of the multimedia interaction image sequence without prior anomaly violation feature data based on the second anomaly violation detection network and the multimedia interaction image sequence without prior anomaly violation feature data;

According to the multimedia interactive image sequence carrying the priori abnormal violation feature data, the multimedia interactive image sequence not carrying the priori abnormal violation feature data and the corresponding first fuzzy abnormal violation feature data, performing differential training on the second abnormal violation detection network to generate a final target abnormal violation detection network;

and detecting the input target multimedia interactive image of the target user based on the final target abnormal violation detection network to obtain abnormal violation segmentation data of the target multimedia interactive image, and correspondingly processing the account of the target user based on the abnormal violation type corresponding to the abnormal violation segmentation data.

In a possible implementation manner of the first aspect, according to the multimedia interactive image sequence carrying the prior abnormal violation feature data, the multimedia interactive image sequence not carrying the prior abnormal violation feature data, and the corresponding first fuzzy abnormal violation feature data, performing differential training on the second abnormal violation detection network to generate a final target abnormal violation detection network, including:

according to the multimedia interactive image sequence carrying the priori abnormal violation feature data, the multimedia interactive image sequence not carrying the priori abnormal violation feature data and the corresponding first fuzzy abnormal violation feature data, performing differential training on the second abnormal violation detection network to generate a third abnormal violation detection network;

Detecting abnormal violation features of the multimedia interactive image sequence which does not carry prior abnormal violation feature data according to the third abnormal violation detection network, and generating second fuzzy abnormal violation feature data of the multimedia interactive image sequence which does not carry prior abnormal violation feature data;

and according to the multimedia interactive image sequence carrying the priori abnormal violation feature data, the multimedia interactive image sequence not carrying the priori abnormal violation feature data and the corresponding second fuzzy abnormal violation feature data, the third abnormal violation detection network carries out differential training to generate a final target abnormal violation detection network.

In a possible implementation manner of the first aspect, generating, based on the second anomaly detection network and the multimedia interactive image sequence that does not carry a priori anomaly violation feature data, first fuzzy anomaly violation feature data of the multimedia interactive image sequence that does not carry a priori anomaly violation feature data includes:

in the initial training process, extracting a fourth number of first unsupervised interactive images from a second number of multimedia interactive image sequences which do not carry priori abnormal violation feature data, wherein the fourth number is not more than the second number;

Detecting abnormal violation features of the first unsupervised interaction image according to the second abnormal violation detection network, and generating first fuzzy abnormal violation feature data and corresponding confidence of the first unsupervised interaction image;

the method comprises the steps of carrying out differential training on the second anomaly violation detection network according to the multimedia interactive image sequence carrying the priori anomaly violation feature data, the multimedia interactive image sequence not carrying the priori anomaly violation feature data and the corresponding first fuzzy anomaly violation feature data, and generating a third anomaly violation detection network, wherein the method comprises the following steps:

based on the confidence level of the first fuzzy anomaly violation feature data, sequencing the first unsupervised interaction image and the corresponding first fuzzy anomaly violation feature data from large to small;

if the total training times from the second abnormal violation detection network to the target abnormal violation detection network are in the training time interval in the initial training process according to the multimedia interactive image sequence carrying the priori abnormal violation feature data and the multimedia interactive image sequence not carrying the priori abnormal violation feature data, selecting a fifth number of first unsupervised interactive images and corresponding first fuzzy abnormal violation feature data before from the fourth number of first unsupervised interactive images and corresponding first fuzzy abnormal violation feature data of the sequencing result;

Selecting the unsupervised interactive images of the current training round and the corresponding first fuzzy abnormal violation feature data according to the size arrangement sequence from the fifth number of first unsupervised interactive images and the corresponding first fuzzy abnormal violation feature data;

selecting a multimedia interactive image sequence carrying the priori abnormal violation characteristic data from the multimedia interactive image sequence carrying the priori abnormal violation characteristic data of the current training round;

and performing differential training on the second abnormal violation detection network according to the unsupervised interaction image of the current training round and the corresponding first fuzzy abnormal violation feature data and the multimedia interaction image sequence of the current training round carrying the priori abnormal violation feature data until the fifth quantity of first unsupervised interaction images and the corresponding first fuzzy abnormal violation feature data are subjected to network dispatching for more than one time, so as to generate the third abnormal violation detection network.

In a possible implementation manner of the first aspect, the detecting the abnormal violation features of the first unsupervised interaction image according to the second abnormal violation detection network, generating first blurred abnormal violation feature data and corresponding confidence degrees of the first unsupervised interaction image includes:

Obtaining a basic fully-connected output unit based on the second abnormal violation detection network;

detecting abnormal violation features of the first unsupervised interaction image according to the basic fully-connected output unit, and generating first fuzzy abnormal violation feature data and corresponding confidence coefficient of the first unsupervised interaction image;

wherein the method further comprises:

and when the second anomaly detection network is subjected to differential training according to the unsupervised interaction image of the current training round, the corresponding first fuzzy anomaly violation feature data and the multimedia interaction image sequence carrying the priori anomaly violation feature data of the current training round, optimizing the basic full-connection output unit based on the second anomaly detection network after the cyclic training until the fifth quantity of first unsupervised interaction images and the corresponding first fuzzy anomaly violation feature data are subjected to network dispatching for more than one time, and generating a first full-connection output unit.

In a possible implementation manner of the first aspect, the method further includes:

increasing the total training times when the second abnormal violation detection network is subjected to differential training in each cycle;

The method for detecting the abnormal violation features of the multimedia interactive image sequence without carrying the priori abnormal violation feature data according to the third abnormal violation detection network comprises the steps of:

if the total training times are smaller than the global training times from the second abnormal violation detection network to the target abnormal violation detection network, extracting a fourth number of second unsupervised interactive images from a second number-fourth number of remaining multimedia interactive image sequences which do not carry priori abnormal violation feature data in the initial training process, wherein the second number is not smaller than twice the fourth number;

performing abnormal violation segmentation data monitoring on the second unsupervised interaction image according to the third abnormal violation detection network to generate second fuzzy abnormal violation feature data and corresponding confidence of the second unsupervised interaction image;

according to the multimedia interactive image sequence carrying the priori abnormal violation feature data, the multimedia interactive image sequence not carrying the priori abnormal violation feature data and the corresponding second fuzzy abnormal violation feature data, the third abnormal violation detection network performs differential training to generate a final target abnormal violation detection network, and the method comprises the following steps:

Based on the confidence level of the second fuzzy anomaly violation feature data, sequencing the second unsupervised interaction image and the corresponding second fuzzy anomaly violation feature data from large to small;

if the total training times are in the training time interval in the initial training process, selecting the fifth number of second supervised interaction images and corresponding second fuzzy anomaly violation feature data from the fourth number of second unsupervised interaction images and corresponding second fuzzy anomaly violation feature data of the sequencing result;

selecting the unsupervised interactive images of the current training round and the corresponding second fuzzy abnormal violation feature data according to the size arrangement sequence from the first fifth number of second unsupervised interactive images and the corresponding second fuzzy abnormal violation feature data;

and iterating the third abnormal violation detection network to carry out differential training according to the unsupervised interaction image of the current training round and the corresponding second fuzzy abnormal violation feature data and the multimedia interaction image sequence carrying the priori abnormal violation feature data of the current training round until the fifth number of second supervised interaction images and the corresponding second fuzzy abnormal violation feature data are adjusted by the network more than once, generating a fourth abnormal violation detection network, and generating the final target abnormal violation detection network.

In a possible implementation manner of the first aspect, the monitoring of the abnormal violation segmentation data of the second unsupervised interaction image according to the third abnormal violation detection network, generating second fuzzy abnormal violation feature data and corresponding confidence degrees of the second unsupervised interaction image, includes:

performing abnormal violation segmentation data monitoring on the second unsupervised interaction image according to the first fully-connected output unit, and generating second fuzzy abnormal violation feature data and corresponding confidence of the second unsupervised interaction image;

wherein the method further comprises:

and when the third anomaly detection network is iterated for differential training according to the unsupervised interaction image of the current training round and the corresponding second fuzzy anomaly violation feature data and the multimedia interaction image sequence carrying the priori anomaly violation feature data of the current training round, optimizing the first fully-connected output unit based on the third anomaly detection network after cyclic training until the sixth number of second supervised interaction images and the corresponding second fuzzy anomaly violation feature data are adjusted by the network more than once, and generating a second fully-connected output unit.

In a possible implementation manner of the first aspect, according to the multimedia interactive image sequence carrying the prior abnormal violation feature data, the multimedia interactive image sequence not carrying the prior abnormal violation feature data, and the corresponding second fuzzy abnormal violation feature data, the third abnormal violation detection network performs differential training to generate the final target abnormal violation detection network, and further includes:

if the total training times are in the training time interval of the backward training process, selecting a sixth previous second unsupervised interactive image and corresponding second fuzzy anomaly violation feature data from the fourth second unsupervised interactive image and corresponding second fuzzy anomaly violation feature data of the sequencing result;

selecting the unsupervised interactive images of the current training round and the corresponding second fuzzy abnormal violation feature data according to the size arrangement sequence from the sixth number of second unsupervised interactive images and the corresponding second fuzzy abnormal violation feature data;

And iterating the third abnormal violation detection network to carry out differential training according to the unsupervised interaction image of the current training round and the corresponding second fuzzy abnormal violation feature data and the multimedia interaction image sequence carrying the priori abnormal violation feature data of the current training round until the sixth number of second unsupervised interaction images and the corresponding second fuzzy abnormal violation feature data are adjusted by the network more than once, generating a fourth abnormal violation detection network, and generating the final target abnormal violation detection network.

In a possible implementation manner of the first aspect, the performing, according to the second anomaly violation detection network, the anomaly violation feature detection on the first unsupervised interaction image, generating a confidence level of first blurred anomaly violation feature data of the first unsupervised interaction image includes:

detecting abnormal violation features of the first unsupervised interaction image according to the second abnormal violation detection network, and generating abnormal violation attributes of image features of each image connected domain;

selecting the abnormal violation attribute of the first image feature of each same image feature as an abnormal violation parameter value;

And obtaining a predicted thermal value of the first fuzzy anomaly violation feature data based on the anomaly violation parameter value, and determining the predicted thermal value as the confidence of the first fuzzy anomaly violation feature data.

detecting abnormal violation features of the first unsupervised interaction image according to the second abnormal violation detection network, and generating first fuzzy abnormal violation feature data and corresponding predicted thermal values of the first unsupervised interaction image;

performing noise feature injection on the first unsupervised interaction image to generate a first enhanced unsupervised interaction image;

detecting abnormal violation segmentation data of the first enhanced unsupervised interaction image according to the second abnormal violation detection network, and generating first enhanced fuzzy abnormal violation feature data and corresponding enhanced prediction thermodynamic values of the first enhanced unsupervised interaction image;

and generating a hit probability value of the first fuzzy anomaly violation feature data based on the first fuzzy anomaly violation feature data and the corresponding predicted thermal value of the first unsupervised interaction image and the first enhanced fuzzy anomaly violation feature data and the corresponding enhanced predicted thermal value of the first enhanced unsupervised interaction image, and determining the hit probability value as the confidence degree of the first fuzzy anomaly violation feature data.

In accordance with one aspect of embodiments of the present application, there is provided a computer device comprising a processor and a machine-readable storage medium having stored therein machine-executable instructions loaded and executed by the processor to implement a RPA-combined multimedia data detection method in any of the foregoing possible implementations.

In accordance with one aspect of embodiments of the present application, there is provided a computer program product or computer program comprising computer instructions stored in a computer readable storage medium. The computer instructions are read from a computer-readable storage medium by a processor of a computer device, which executes the computer instructions, causing the computer device to perform the methods provided in the various alternative implementations of the three aspects described above.

According to an aspect of the embodiment of the present application, there is provided a multimedia data detection apparatus in combination with RPA, implemented by a computer device running an automated RPA system with a robot flow, the apparatus comprising:

the acquisition module is used for acquiring a first number of multimedia interaction image sequences carrying priori abnormal violation feature data and a second number of multimedia interaction image sequences not carrying priori abnormal violation feature data;

The first generation module is used for carrying out unsupervised network learning optimization on the basic abnormal violation detection network according to the multimedia interactive image sequence which does not carry priori abnormal violation feature data, and generating a first abnormal violation detection network;

the second generation module is used for carrying out network learning optimization on the first abnormal violation detection network traversal for a third number of times according to the multimedia interactive image sequence carrying the priori abnormal violation feature data, so as to generate a second abnormal violation detection network;

the third generation module is used for generating first fuzzy anomaly violation feature data of the multimedia interaction image sequence without the prior anomaly violation feature data based on the second anomaly violation detection network and the multimedia interaction image sequence without the prior anomaly violation feature data;

the fourth generation module is used for carrying out differential training on the second abnormal violation detection network according to the multimedia interactive image sequence carrying the priori abnormal violation feature data, the multimedia interactive image sequence not carrying the priori abnormal violation feature data and the corresponding first fuzzy abnormal violation feature data, so as to generate a final target abnormal violation detection network;

And the account processing module is used for detecting the target multimedia interaction image of the target user based on the final target abnormal violation detection network, obtaining abnormal violation segmentation data of the target multimedia interaction image, and correspondingly processing the account of the target user based on the abnormal violation type corresponding to the abnormal violation segmentation data.

In the technical schemes provided by some embodiments of the present application, when a first number of multimedia interactive image sequences carrying priori abnormal violation feature data and a second number of multimedia interactive image sequences not carrying priori abnormal violation feature data are obtained, unsupervised network learning optimization is performed on a basic abnormal violation detection network according to the multimedia interactive image sequences not carrying the priori abnormal violation feature data, so as to generate a final first abnormal violation detection network; the first abnormal violation detection network is trained circularly for a third number of times according to the multimedia interactive image sequence carrying the priori abnormal violation feature data, and a final second abnormal violation detection network is generated; generating first fuzzy anomaly violation feature data of the multimedia interaction image sequence without prior anomaly violation feature data based on the second anomaly violation detection network and the multimedia interaction image sequence without prior anomaly violation feature data; and then combining the multimedia interactive image sequence carrying the priori abnormal violation feature data with the multimedia interactive image sequence not carrying the priori abnormal violation feature data and the corresponding first fuzzy abnormal violation feature data to train the second abnormal violation detection network, thereby improving the training speed of the abnormal violation detection network and reducing the workload of training labeling.

Drawings

For a clearer description of the technical solutions of the embodiments of the present application, reference will be made to the accompanying drawings, which are needed to be activated in the embodiments, and it should be understood that the following drawings only illustrate some embodiments of the present application, and therefore should not be considered as limiting the scope, and other related drawings can be extracted by those skilled in the art without the inventive effort.

Fig. 1 is a flow chart of a method for detecting multimedia data in combination with RPA according to an embodiment of the present application;

fig. 2 is a schematic block diagram of a module assembly of an RPA-combined multimedia data detection device according to an embodiment of the present application;

fig. 3 is a schematic block diagram of a computer device for implementing the method for detecting multimedia data in combination with RPA according to an embodiment of the present application.

Detailed Description

The following description is presented to enable one of ordinary skill in the art to make and use the application and is provided in the context of a particular application and its requirements. It will be apparent to those having ordinary skill in the art that various changes can be made to the disclosed embodiments and that the general principles defined herein may be applied to other embodiments and applications without departing from the principles and scope of the application. Therefore, the present application is not limited to the described embodiments, but is to be accorded the widest scope consistent with the claims.

Fig. 1 is a flowchart of an RPA-combined multimedia data detection method according to an embodiment of the present application, and the RPA-combined multimedia data detection method is described in detail below.

In step S110, a first number of multimedia interactive image sequences W carrying a priori abnormal violation feature data and a second number of multimedia interactive image sequences G not carrying a priori abnormal violation feature data are obtained.

In an alternative embodiment, the multimedia interaction image sequence W carrying the prior abnormal violation feature data includes an initial multimedia interaction image sample and corresponding prior abnormal violation feature data, where the prior abnormal violation feature data may be manually annotated manually based on prior knowledge and determined as a training tag of the corresponding initial multimedia interaction image sample. The multimedia interaction image sequence G, which does not carry a priori abnormal violation feature data, comprises an initial multimedia interaction image sample, but does not comprise corresponding a priori abnormal violation feature data. The embodiment of the application does not limit the data sources of the multimedia interactive image sequence carrying the priori abnormal violation characteristic data and the multimedia interactive image sequence not carrying the priori abnormal violation characteristic data, and only needs to acquire under the authorization permission of the related user.

In this embodiment, the abnormal violation feature data may refer to specific image area data in which abnormal violations exist in the multimedia interactive image, such as image area data related to rules that violate legal regulations, multimedia platform rules, ethical rules, and the like.

In an alternative embodiment, the second number may be not smaller than the first number, that is, the number of multimedia interactive image sequences that do not carry a priori abnormal violation feature data may be greater than the number of multimedia interactive image sequences that carry a priori abnormal violation feature data, so that the labeling time of the initial multimedia interactive image sample may be reduced, and the collection cost of the multimedia interactive image sequences that carry a priori abnormal violation feature data may be reduced.

In step S120, an unsupervised network learning optimization is performed on the basic anomaly detection network according to the multimedia interactive image sequence that does not carry the prior anomaly violation feature data, so as to generate a first anomaly violation detection network.

In an alternative embodiment, the base anomaly violation detection network is first subjected to unsupervised network learning optimization according to a multimedia interactive image sequence G which does not carry priori anomaly violation feature data, and a first anomaly violation detection network is generated.

In an alternative implementation manner, the embodiment of the present application may not limit the specific algorithm structure of the abnormal violation detection network, for example, a convolutional neural network model may be adopted. When the network is detected according to different abnormal violations, corresponding transformation is performed according to the non-supervision training scheme. In the embodiment of the application, an anomaly violation detection network which is not trained is used as a basic anomaly violation detection network, the basic anomaly violation detection network which is not supervised network learning optimization is used as a first anomaly violation detection network according to a multimedia interactive image sequence which does not carry priori anomaly violation feature data, the first anomaly violation detection network which is circularly trained for a third number of times according to the multimedia interactive image sequence which carries the priori anomaly violation feature data is used as a second anomaly violation detection network, and then the weight parameters of the second anomaly violation detection network can be updated by combining the multimedia interactive image sequence which carries the priori anomaly violation feature data and the multimedia interactive image sequence which does not carry the priori anomaly violation feature data and the corresponding first fuzzy anomaly violation feature data at the same time, so that a final target anomaly violation detection network is generated.

In step S130, performing network learning optimization on the first abnormal violation detection network traversal for a third number of times according to the multimedia interactive image sequence W carrying the prior abnormal violation feature data, so as to generate a second abnormal violation detection network.

In an alternative embodiment, in the process of continuously updating the first abnormal violation detection network after the unsupervised network learning optimization in the step S120, the first abnormal violation detection network may be initially trained according to the multimedia interactive image sequence W carrying the prior abnormal violation feature data only in a circulating manner, so as to ensure that the second abnormal violation detection network has a certain distinguishing performance of the fuzzy abnormal violation feature data when distinguishing the fuzzy abnormal violation feature data of the multimedia interactive image sequence G not carrying the prior abnormal violation feature data in the step S140 described below.

In an alternative embodiment, the third number is smaller than the set number, where the set number refers to the number of times of pre-estimation required for training the first abnormal violation detection network to converge according to the multimedia interactive image sequence W carrying the prior abnormal violation feature data, that is, when the first abnormal violation detection network is traversed according to the multimedia interactive image sequence W carrying the prior abnormal violation feature data for performing network learning and optimizing the third number of times, the first abnormal violation detection network is converged without training, that is, the first third number step only needs to enable the second abnormal violation detection network to have the preliminary discrimination performance of fuzzy abnormal violation feature data, and then the non-supervision learning process of the multimedia interactive image sequence W carrying the prior abnormal violation feature data and the multimedia interactive image sequence G not carrying the prior abnormal violation feature data in step S150 can be performed, so that the first abnormal violation detection network converges.

In step S140, based on the second abnormal violation detection network and the multimedia interactive image sequence that does not carry a priori abnormal violation feature data, first fuzzy abnormal violation feature data of the multimedia interactive image sequence that does not carry a priori abnormal violation feature data is generated.

The fuzzy anomaly violation feature data may include first fuzzy anomaly violation feature data, second fuzzy anomaly violation feature data, and the like, for distinguishing fuzzy anomaly violation feature data generated by different image connected domains. In an alternative embodiment, the second anomaly detection network may be generated according to the multimedia interactive image sequence W carrying the prior anomaly violation feature data, and the detection may be performed on the multimedia interactive image sequence G not carrying the prior anomaly violation feature data, so as to obtain the first fuzzy anomaly violation feature data of the multimedia interactive image sequence G not carrying the prior anomaly violation feature data, and then the multimedia interactive image sequence W carrying the prior anomaly violation feature data and the multimedia interactive image sequence G not carrying the prior anomaly violation feature data and the corresponding first fuzzy anomaly violation feature data are input into the second anomaly detection network again for updating.

In step S150, according to the multimedia interactive image sequence carrying the prior abnormal violation feature data, the multimedia interactive image sequence not carrying the prior abnormal violation feature data, and the corresponding first fuzzy abnormal violation feature data, the second abnormal violation detection network is trained in a distinguishing manner, so as to generate a final target abnormal violation detection network.

In step S160, the final target abnormal violation detection network is used to detect the target multimedia interaction image of the target user, obtain abnormal violation segmentation data of the target multimedia interaction image, and perform corresponding processing on the account of the target user based on the abnormal violation type corresponding to the abnormal violation segmentation data.

For example, after the abnormal violation segmentation data is obtained, the corresponding abnormal violation type can be determined, and then the account of the target user is correspondingly processed based on account processing rules corresponding to different abnormal violation types.

In an alternative embodiment, after the first fuzzy anomaly violation feature data of the multimedia interactive image sequence without the prior anomaly violation feature data is obtained according to step S140, the image data is extracted from the multimedia interactive image sequence W with the prior anomaly violation feature data and the multimedia interactive image sequence without the prior anomaly violation feature data of the first fuzzy anomaly violation feature data to update the weight parameters of the second anomaly violation detection network, so that the target anomaly violation detection network is obtained, the weight parameters of the second anomaly violation detection network are not required to be reset, and the target anomaly violation detection network can be generated without training again.

In an alternative embodiment, the performing the differential training on the second anomaly detection network according to the multimedia interactive image sequence carrying the prior anomaly violation feature data, the multimedia interactive image sequence not carrying the prior anomaly violation feature data, and the corresponding first fuzzy anomaly violation feature data, to generate a final target anomaly violation detection network may include: according to the multimedia interactive image sequence carrying the priori abnormal violation feature data, the multimedia interactive image sequence not carrying the priori abnormal violation feature data and the corresponding first fuzzy abnormal violation feature data, performing differential training on the second abnormal violation detection network to generate a third abnormal violation detection network; detecting abnormal violation features of the multimedia interactive image sequence which does not carry prior abnormal violation feature data according to the third abnormal violation detection network, and generating second fuzzy abnormal violation feature data of the multimedia interactive image sequence which does not carry prior abnormal violation feature data; and according to the multimedia interactive image sequence carrying the priori abnormal violation feature data, the multimedia interactive image sequence not carrying the priori abnormal violation feature data and the corresponding second fuzzy abnormal violation feature data, the third abnormal violation detection network carries out differential training to generate a final target abnormal violation detection network.

In an alternative embodiment, in the process from the second abnormal violation detection network obtaining to the final target abnormal violation detection network obtaining, fuzzy abnormal violation feature data of the multimedia interactive image sequence G without carrying prior abnormal violation feature data are continuously generated and continuously changed, instead of directly generating all fuzzy abnormal violation feature data of the multimedia interactive image sequence G without carrying prior abnormal violation feature data based on the updated second abnormal violation detection network, then updating the second abnormal violation detection network according to the multimedia interactive image sequence G with the prior abnormal violation feature data and all fuzzy abnormal violation feature data of the multimedia interactive image sequence G without carrying the prior abnormal violation feature data, in the updating scheme, the fuzzy abnormal feature data of the multimedia interactive image sequence G without carrying the prior abnormal violation feature data are continuously changed along with training, and in order to accurately predict the third abnormal violation network detection according to the third abnormal violation network detection by the second abnormal violation detection network, for example, the third abnormal violation network detection is more accurate than the third abnormal network detection by the third abnormal violation network 35, and the third abnormal violation network detection is more accurate than the third abnormal network detection by the third abnormal violation network detection network 35, as the second fuzzy anomaly violation feature data), the accuracy may be better than the fuzzy anomaly violation feature data (as the first fuzzy anomaly violation feature data) predicted by the second anomaly violation detection network, the accuracy of the fuzzy anomaly violation feature data (as the third fuzzy anomaly violation feature data) predicted by the fourth anomaly violation detection network may be higher than the accuracy of the fuzzy anomaly violation feature data (i.e., the second fuzzy anomaly violation feature data) predicted by the third anomaly violation detection network, the accuracy of …, i.e., the fuzzy anomaly violation feature data may be gradually improved, and as the accuracy of the fuzzy anomaly violation feature data gradually improves, the accuracy of the anomaly violation detection network (e.g., the third anomaly violation detection network, the fourth anomaly violation detection network, etc.) generated by updating the fuzzy anomaly violation feature data may be further improved, resulting in that the accuracy of the final updated target anomaly violation detection network is greatly optimized.

In an alternative embodiment, based on the second anomaly detection network and the multimedia interactive image sequence that does not carry a priori anomaly violation feature data, generating the first fuzzy anomaly violation feature data of the multimedia interactive image sequence that does not carry a priori anomaly violation feature data may include: in the initial training process, extracting a fourth number of first unsupervised interactive images from a second number of multimedia interactive image sequences which do not carry priori abnormal violation feature data, wherein the fourth number is not more than the second number; and detecting abnormal violation characteristics of the first unsupervised interaction image according to the second abnormal violation detection network, and generating first fuzzy abnormal violation characteristic data and corresponding confidence of the first unsupervised interaction image.

The step of performing differential training on the second anomaly violation detection network according to the multimedia interactive image sequence carrying the priori anomaly violation feature data, the multimedia interactive image sequence not carrying the priori anomaly violation feature data and the corresponding first fuzzy anomaly violation feature data to generate a third anomaly violation detection network may include: based on the confidence level of the first fuzzy anomaly violation feature data, sequencing the first unsupervised interaction image and the corresponding first fuzzy anomaly violation feature data from large to small; if the total training times from the second abnormal violation detection network to the target abnormal violation detection network are in the training time interval in the initial training process according to the multimedia interactive image sequence carrying the priori abnormal violation feature data and the multimedia interactive image sequence not carrying the priori abnormal violation feature data, selecting a fifth number of first unsupervised interactive images and corresponding first fuzzy abnormal violation feature data before from the fourth number of first unsupervised interactive images and corresponding first fuzzy abnormal violation feature data of the sequencing result; selecting the unsupervised interactive images of the current training round and the corresponding first fuzzy abnormal violation feature data according to the size arrangement sequence from the fifth number of first unsupervised interactive images and the corresponding first fuzzy abnormal violation feature data; selecting a multimedia interactive image sequence carrying the priori abnormal violation characteristic data from the multimedia interactive image sequence carrying the priori abnormal violation characteristic data of the current training round; and performing differential training on the second abnormal violation detection network according to the unsupervised interaction image of the current training round and the corresponding first fuzzy abnormal violation feature data and the multimedia interaction image sequence of the current training round carrying the priori abnormal violation feature data, until the fifth quantity of first unsupervised interaction images and the corresponding first fuzzy abnormal violation feature data are adjusted by the network more than once, and generating the third abnormal violation detection network.

In an alternative embodiment, the weight parameter updating of the second abnormal violation detection network to the target abnormal violation detection network may be performed in a distributed manner in the whole process according to the multimedia interactive image sequence carrying the priori abnormal violation feature data, the multimedia interactive image sequence not carrying the priori abnormal violation feature data, and the corresponding fuzzy abnormal violation feature data.

For example, in the initial training process, when the fourth number is equal to the second number, the second number of multimedia interactive image sequences without prior abnormal violation feature data is used as the second number of first unsupervised interactive images, and the second number of first fuzzy abnormal violation feature data and the corresponding confidence level of the second number of multimedia interactive image sequences without prior abnormal violation feature data are respectively obtained by prediction.

When the fourth number is smaller than the second number, a temporary fuzzy anomaly violation feature database can be established, a fourth number of multimedia interaction image sequences which do not carry prior anomaly violation feature data are extracted from the second number of multimedia interaction image sequences which do not carry prior anomaly violation feature data, namely a fourth number of first unsupervised interaction images are obtained, the first fuzzy anomaly violation feature data and the corresponding confidence of the extracted fourth number of first unsupervised interaction images are predicted according to the second anomaly violation detection network, and the fourth number of first unsupervised interaction images and the corresponding first fuzzy anomaly violation feature data are added into the temporary fuzzy anomaly violation feature database.

In the process of updating the second abnormal violation detection network, according to the confidence level of the first fuzzy abnormal violation feature data, selecting a first unsupervised interactive image of a first fifth number and corresponding first fuzzy abnormal violation feature data from the temporary fuzzy abnormal violation feature database, discarding other remaining first unsupervised interactive images and corresponding first fuzzy abnormal violation feature data in the temporary fuzzy abnormal violation feature database, then sequentially extracting the unsupervised interactive image of the current training round and corresponding first fuzzy abnormal violation feature data according to the confidence level of the first fuzzy abnormal violation feature data from the first unsupervised interactive image of the first fifth number and corresponding first fuzzy abnormal violation feature data, and simultaneously combining a multimedia interactive image sequence carrying the priori abnormal violation feature data of the current training round extracted from a multimedia interactive image sequence carrying the priori abnormal violation feature data, continuously circularly updating the weight parameters of the second abnormal violation detection network until the first unsupervised interactive image of the first fifth number and corresponding first fuzzy abnormal violation feature data are sequentially extracted, and the first fuzzy abnormal violation feature data of the first abnormal violation detection network is detected as the first abnormal illegal network.

And extracting a fourth multimedia interactive image sequence which does not carry priori abnormal violation feature data from the rest second quantity-fourth multimedia interactive image sequences which do not carry priori abnormal violation feature data, taking the fourth multimedia interactive image sequence as a fourth second unsupervised interactive image, predicting second fuzzy abnormal violation feature data and corresponding confidence of the fourth second unsupervised interactive image according to a third abnormal violation detection network, and putting the fourth second unsupervised interactive image and the corresponding second fuzzy abnormal violation feature data into the temporary fuzzy abnormal violation feature database, so that circulation is performed until the maximum updating times are reached, and generating a target abnormal violation detection network.

And when the fourth number is smaller than the global training times of each stage, the fuzzy abnormal violation characteristic data of the multimedia interactive image sequence without carrying the priori abnormal violation characteristic data can be updated for multiple rounds.

In an alternative embodiment, the first extracted fourth number of multimedia interactive image sequences without the a priori abnormal violation feature data is used as the first unsupervised interactive image, and the like, and the kth extracted fourth number of multimedia interactive image sequences without the a priori abnormal violation feature data is used as the kth unsupervised interactive image.

In an alternative embodiment, the unsupervised network learning is performed according to the multimedia interactive image sequence without the prior abnormal violation feature data, the second abnormal violation detection network after training the third number of times of the multimedia interactive image sequence with the prior abnormal violation feature data carries out abnormal violation detection on the fourth number of first unsupervised interactive images, so as to obtain first fuzzy abnormal violation feature data of the fourth number of first unsupervised interactive images, along with the first fuzzy abnormal violation feature data, the second abnormal violation detection network can simultaneously generate a confidence level describing the abnormal violation probability of the first fuzzy abnormal violation feature data, and the confidence level can reflect the accuracy of the first fuzzy abnormal violation feature data generated by the second abnormal violation detection network. When the first unsupervised interaction image and the corresponding first fuzzy anomaly violation feature data used for circularly updating the weight parameters of the second anomaly violation detection network are selected from the fourth number of first unsupervised interaction images and the corresponding first fuzzy anomaly violation feature data of the sequencing result, the first unsupervised interaction image and the corresponding first fuzzy anomaly violation feature data with relatively higher confidence level, such as the first unsupervised interaction image and the corresponding first fuzzy anomaly violation feature data of the fifth number, can be selected so as to improve the accuracy of the third anomaly violation detection network generated by updating, thereby improving the accuracy of the target anomaly violation detection network.

In an alternative embodiment, in the process of updating the weight parameters of the second abnormal violation detection network according to the fourth number of first unsupervised interaction images and the corresponding first fuzzy abnormal violation feature data, the first unsupervised interaction images and the corresponding first fuzzy abnormal violation feature data, which are selected to update the weight parameters of the second abnormal violation detection network, may be determined based on the confidence level of the first fuzzy abnormal violation feature data of the first unsupervised interaction images, so as to avoid the situation that the subsequent detection accuracy is lower due to training of randomly selecting the multimedia interaction image sequence without the prior abnormal violation feature data and the corresponding fuzzy abnormal violation feature data, because the quality of the fuzzy abnormal violation feature data of the multimedia interaction image sequence without the prior abnormal violation feature data is different, and randomly selecting the multimedia interaction image sequence without the prior abnormal violation feature data and the corresponding fuzzy abnormal violation feature data, which may cause quality crossover, are used for subsequent network learning.

In an alternative embodiment, the detecting the abnormal violation feature of the first unsupervised interaction image according to the second abnormal violation detection network, and generating the first fuzzy abnormal violation feature data and the corresponding confidence coefficient of the first unsupervised interaction image may include: obtaining a basic fully-connected output unit based on the second abnormal violation detection network; and detecting abnormal violation characteristics of the first unsupervised interaction image according to the basic fully-connected output unit, and generating first fuzzy abnormal violation characteristic data and corresponding confidence of the first unsupervised interaction image.

Wherein the method may further comprise: and when the second anomaly detection network is subjected to differential training according to the unsupervised interaction image of the current training round and the corresponding first fuzzy anomaly violation feature data and the multimedia interaction image sequence carrying the priori anomaly violation feature data of the current training round, optimizing the basic full-connection output unit based on the second anomaly detection network after the cyclic training until the fifth quantity of first unsupervised interaction images and the corresponding first fuzzy anomaly violation feature data are subjected to network dispatching for more than one time, and generating a first full-connection output unit.

In an alternative embodiment, the first fuzzy anomaly violation feature data and the corresponding confidence level of the first unsupervised interaction image may be predicted directly from the second anomaly violation detection network.

In an alternative embodiment, in the processing process of the first unsupervised interactive images and the corresponding first fuzzy anomaly violation feature data of the fifth number, the unsupervised interactive images and the corresponding first fuzzy anomaly violation feature data of the current training round can be determined from the first unsupervised interactive images and the corresponding first fuzzy anomaly violation feature data of the fifth number, and the second anomaly detection network is circularly updated with the weight parameters according to the unsupervised interactive images and the corresponding first fuzzy anomaly violation feature data of the current training round and combined with the multimedia interactive image sequence carrying the priori anomaly violation feature data of the current training round until the first unsupervised interactive images and the corresponding first fuzzy anomaly violation feature data of the fifth number are all scheduled by the network for more than one time, and then a third anomaly detection network is generated.

According to the unsupervised interactive image of the first current training round, the corresponding first fuzzy abnormal violation feature data and the multimedia interactive image sequence training iteration carrying the priori abnormal violation feature data of the first current training round, a second abnormal violation detection network after the first cyclic training is obtained, and then a basic fully-connected output unit is updated according to the second abnormal violation detection network after the first cyclic training, so that a first fully-connected output unit is generated; and then training and iterating according to the unsupervised interactive image of the second current training round and the corresponding first fuzzy anomaly violation feature data and the multimedia interactive image sequence carrying the priori anomaly violation feature data of the second current training round to obtain a second anomaly violation detection network after the second cyclic training, updating the first full-connection output unit according to the second anomaly violation detection network after the second cyclic training, generating updated first full-connection output units … and so on until the fifth quantity of first unsupervised interactive images and the corresponding first fuzzy anomaly violation feature data are at least once according to each iteration of the second anomaly violation detection network.

In an alternative implementation, the above embodiment may further include: the total number of training is incremented as the second anomaly violation detection network is differentially trained for each cycle.

The step of detecting the abnormal violation features of the multimedia interactive image sequence without the prior abnormal violation feature data according to the third abnormal violation detection network, and the step of generating the second fuzzy abnormal violation feature data of the multimedia interactive image sequence without the prior abnormal violation feature data may include: if the total training times are smaller than the global training times from the second abnormal violation detection network to the target abnormal violation detection network, extracting a fourth number of second unsupervised interaction images from a second number-fourth number of remaining multimedia interaction image sequences which do not carry priori abnormal violation feature data in the initial training process; and monitoring the abnormal violation segmentation data of the second unsupervised interaction image according to the third abnormal violation detection network, and generating second fuzzy abnormal violation feature data and corresponding confidence of the second unsupervised interaction image.

According to the multimedia interactive image sequence carrying the priori abnormal violation feature data, the multimedia interactive image sequence not carrying the priori abnormal violation feature data and the corresponding second fuzzy abnormal violation feature data, the third abnormal violation detection network performs differential training to generate a final target abnormal violation detection network, which may include: based on the confidence level of the second fuzzy anomaly violation feature data, sequencing the second unsupervised interaction image and the corresponding second fuzzy anomaly violation feature data from large to small; if the total training times are in the training time interval in the initial training process, selecting the fifth number of second unsupervised interaction images and corresponding second fuzzy anomaly violation feature data from the fourth number of second unsupervised interaction images and corresponding second fuzzy anomaly violation feature data of the sequencing result; selecting the unsupervised interactive images of the current training round and the corresponding second fuzzy abnormal violation feature data according to the size arrangement sequence from the first fifth number of second unsupervised interactive images and the corresponding second fuzzy abnormal violation feature data; selecting a multimedia interactive image sequence carrying the priori abnormal violation characteristic data from the multimedia interactive image sequence carrying the priori abnormal violation characteristic data of the current training round; and iterating the third abnormal violation detection network to carry out differential training according to the unsupervised interaction image of the current training round and the corresponding second fuzzy abnormal violation feature data and the multimedia interaction image sequence carrying the priori abnormal violation feature data of the current training round until the fifth number of second unsupervised interaction images and the corresponding second fuzzy abnormal violation feature data are adjusted by the network more than once, generating a fourth abnormal violation detection network, and generating the final target abnormal violation detection network.

In an alternative embodiment, the monitoring of the abnormal violation segmentation data of the second unsupervised interaction image according to the third abnormal violation detection network, and generating the second fuzzy abnormal violation feature data and the corresponding confidence coefficient of the second unsupervised interaction image may include: and monitoring the abnormal violation segmentation data of the second unsupervised interaction image according to the first fully-connected output unit, and generating second fuzzy abnormal violation feature data and corresponding confidence of the second unsupervised interaction image.

The above embodiment may further include: and when the third anomaly detection network is iterated for differential training according to the unsupervised interaction image of the current training round and the corresponding second fuzzy anomaly violation feature data and the multimedia interaction image sequence carrying the priori anomaly violation feature data of the current training round, optimizing the first fully-connected output unit based on the third anomaly detection network after cyclic training until the sixth number of second unsupervised interaction images and the corresponding second fuzzy anomaly violation feature data are subjected to network dispatching for more than one time, and generating a second fully-connected output unit.

In an alternative embodiment, according to the multimedia interactive image sequence carrying the prior abnormal violation feature data, the multimedia interactive image sequence not carrying the prior abnormal violation feature data, and the corresponding second fuzzy abnormal violation feature data, the third abnormal violation detection network performs differential training to generate the final target abnormal violation detection network, and may further include: if the total training times are in the training time interval of the backward training process, selecting a sixth previous second unsupervised interactive image and corresponding second fuzzy anomaly violation feature data from the fourth second unsupervised interactive image and corresponding second fuzzy anomaly violation feature data of the sequencing result; selecting the unsupervised interactive images of the current training round and the corresponding second fuzzy abnormal violation feature data according to the size arrangement sequence from the sixth number of second unsupervised interactive images and the corresponding second fuzzy abnormal violation feature data; selecting a multimedia interactive image sequence carrying the priori abnormal violation characteristic data from the multimedia interactive image sequence carrying the priori abnormal violation characteristic data of the current training round; and iterating the third abnormal violation detection network to carry out differential training according to the unsupervised interaction image of the current training round and the corresponding second fuzzy abnormal violation feature data and the multimedia interaction image sequence carrying the priori abnormal violation feature data of the current training round until the sixth number of second unsupervised interaction images and the corresponding second fuzzy abnormal violation feature data are adjusted by the network more than once, generating a fourth abnormal violation detection network, and generating the final target abnormal violation detection network.

In an alternative embodiment, the detecting the abnormal violation feature of the first unsupervised interaction image according to the second abnormal violation detection network, and generating the confidence level of the first fuzzy abnormal violation feature data of the first unsupervised interaction image may include: detecting abnormal violation features of the first unsupervised interaction image according to the second abnormal violation detection network, and generating abnormal violation attributes of image features of each image connected domain; selecting the abnormal violation attribute of the first image feature of each same image feature as an abnormal violation parameter value; and obtaining a predicted thermal value of the first fuzzy anomaly violation feature data based on the anomaly violation parameter value, and determining the predicted thermal value as the confidence of the first fuzzy anomaly violation feature data.

In an alternative embodiment, the detecting the abnormal violation feature of the first unsupervised interaction image according to the second abnormal violation detection network, and generating the confidence level of the first fuzzy abnormal violation feature data of the first unsupervised interaction image may include: detecting abnormal violation features of the first unsupervised interaction image according to the second abnormal violation detection network, and generating first fuzzy abnormal violation feature data and corresponding predicted thermal values of the first unsupervised interaction image; performing noise feature injection on the first unsupervised interaction image to generate a first enhanced unsupervised interaction image; detecting abnormal violation segmentation data of the first enhanced unsupervised interaction image according to the second abnormal violation detection network, and generating first enhanced fuzzy abnormal violation feature data and corresponding enhanced prediction thermodynamic values of the first enhanced unsupervised interaction image; and generating a hit probability value of the first fuzzy anomaly violation feature data based on the first fuzzy anomaly violation feature data and the corresponding predicted thermal value of the first unsupervised interaction image and the first enhanced fuzzy anomaly violation feature data and the corresponding enhanced predicted thermal value of the first enhanced unsupervised interaction image, and determining the hit probability value as the confidence degree of the first fuzzy anomaly violation feature data.

In an alternative embodiment, the first unsupervised interaction image may be enhanced once (i.e. the noise feature injection is performed), so as to generate a first enhanced unsupervised interaction image, and the first fuzzy anomaly violation feature data of the first enhanced unsupervised interaction image may be re-predicted according to the second anomaly detection network or the basic fully-connected output unit, as the first enhanced fuzzy anomaly violation feature data of the first enhanced unsupervised interaction image, while the fourth number of the first enhanced fuzzy anomaly violation feature data is predicted, as the enhanced prediction thermal value, and the fourth number R third number of the first fuzzy anomaly violation feature data of the first unsupervised interaction image may be generated according to the fourth number of the first fuzzy anomaly violation feature data of the first unsupervised interaction image and the enhanced prediction thermal value of the first enhanced unsupervised interaction image.

Based on the steps, when a first number of multimedia interactive image sequences carrying priori abnormal violation feature data and a second number of multimedia interactive image sequences not carrying priori abnormal violation feature data are obtained, performing unsupervised network learning optimization on a basic abnormal violation detection network according to the multimedia interactive image sequences not carrying the priori abnormal violation feature data, and generating a final first abnormal violation detection network; the first abnormal violation detection network is trained circularly for a third number of times according to the multimedia interactive image sequence carrying the priori abnormal violation feature data, and a final second abnormal violation detection network is generated; generating first fuzzy anomaly violation feature data of the multimedia interaction image sequence without prior anomaly violation feature data based on the second anomaly violation detection network and the multimedia interaction image sequence without prior anomaly violation feature data; and then combining the multimedia interactive image sequence carrying the priori abnormal violation feature data with the multimedia interactive image sequence not carrying the priori abnormal violation feature data and the corresponding first fuzzy abnormal violation feature data to train the second abnormal violation detection network, thereby improving the training speed of the abnormal violation detection network and reducing the workload of training labeling.

Another training embodiment of the above anomaly violation detection network is provided below.

In step S201, an unsupervised network learning optimization is performed on the basic anomaly detection network according to a multimedia interactive image sequence that does not carry a priori anomaly violation feature data, so as to generate a first anomaly violation detection network.

In step S202, network learning is performed on the first abnormal violation detection network traversal for a third number of times according to the multimedia interactive image sequence carrying the priori abnormal violation feature data, so as to generate a second abnormal violation detection network.

In step S203, the abnormal violation feature detection is performed on the multimedia interactive image sequence without the prior abnormal violation feature data according to the second abnormal violation detection network, so as to generate first fuzzy abnormal violation feature data of the multimedia interactive image sequence without the prior abnormal violation feature data, that is, the first fuzzy abnormal violation feature data multimedia interactive image sequence includes the multimedia interactive image sequence without the prior abnormal violation feature data and the corresponding first fuzzy abnormal violation feature data.

In step S204, the weight parameter of the second abnormal violation detection network is updated by combining the multimedia interactive image sequence carrying the prior abnormal violation feature data and the multimedia interactive image sequence of the first fuzzy abnormal violation feature data together, so as to generate a third abnormal violation detection network.

In step S205, abnormal violation feature detection is performed on the multimedia interactive image sequence without the prior abnormal violation feature data according to the third abnormal violation detection network, so as to generate second fuzzy abnormal violation feature data of the multimedia interactive image sequence without the prior abnormal violation feature data, that is, the second fuzzy abnormal violation feature data multimedia interactive image sequence includes the multimedia interactive image sequence without the prior abnormal violation feature data and the corresponding second fuzzy abnormal violation feature data.

In step S206, the third abnormal violation detection network is trained by combining the multimedia interactive image sequence carrying the prior abnormal violation feature data and the multimedia interactive image sequence of the second fuzzy abnormal violation feature data, a fourth abnormal violation detection network is generated, and the like, and finally the target abnormal violation detection network is obtained.

Fig. 2 is a functional block diagram of an RPA-combined multimedia data detection device 200 according to an embodiment of the present application, where functions implemented by the RPA-combined multimedia data detection device 200 may correspond to steps executed by the above-described method. The RPA-combined multimedia data detection apparatus 200 may be understood as the above-mentioned computer device, or a processor of the computer device, or may be understood as a component, which is independent from the above-mentioned computer device or processor and is controlled by the computer device, to implement the functions of the present application, as shown in fig. 2, and the functions of the respective functional modules of the RPA-combined multimedia data detection apparatus 200 are described in detail below.

An obtaining module 210, configured to obtain a first number of multimedia interactive image sequences that carry a priori abnormal violation feature data and a second number of multimedia interactive image sequences that do not carry a priori abnormal violation feature data;

the first generating module 220 is configured to perform unsupervised network learning optimization on the basic anomaly detection network according to the multimedia interactive image sequence that does not carry the prior anomaly violation feature data, so as to generate a first anomaly violation detection network;

the second generating module 230 is configured to perform network learning optimization on the first abnormal violation detection network traversal for a third number of times according to the multimedia interactive image sequence carrying the prior abnormal violation feature data, so as to generate a second abnormal violation detection network;

a third generating module 240, configured to generate, based on the second abnormal violation detection network and the multimedia interactive image sequence that does not carry a priori abnormal violation feature data, first fuzzy abnormal violation feature data of the multimedia interactive image sequence that does not carry a priori abnormal violation feature data;

a fourth generating module 250, configured to perform differential training on the second abnormal violation detection network according to the multimedia interactive image sequence carrying the prior abnormal violation feature data, the multimedia interactive image sequence not carrying the prior abnormal violation feature data, and the corresponding first fuzzy abnormal violation feature data, so as to generate a final target abnormal violation detection network;

The account processing module 260 is configured to detect the input target multimedia interaction image of the target user based on the final target abnormal violation detection network, obtain abnormal violation segmentation data of the target multimedia interaction image, and perform corresponding processing on the account of the target user based on an abnormal violation type corresponding to the abnormal violation segmentation data.

Fig. 3 illustrates a hardware structural view of a computer device 100 for implementing the RPA-combined multimedia data detection method according to an embodiment of the present application, and as shown in fig. 3, the computer device 100 may include a processor 110, a machine-readable storage medium 120, a bus 130, and a communication unit 140.

In some alternative embodiments, the computer device 100 may be a single server or a group of servers. The server farm may be centralized or distributed (e.g., computer device 100 may be a distributed system). In some alternative embodiments, the computer device 100 may be local or remote. For example, computer device 100 may access information and/or data stored in machine-readable storage medium 120 via a network. As another example, computer device 100 may be directly connected to machine-readable storage medium 120 to access stored information and/or data. In some alternative implementations, the computer device 100 may be implemented on a cloud platform. For example only, the cloud platform may include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, or the like, or any combination thereof.

The machine-readable storage medium 120 may store data and/or instructions. In some alternative implementations, the machine-readable storage medium 120 may store data acquired from an external terminal. In some alternative implementations, the machine-readable storage medium 120 may store data and/or instructions that are used by the computer device 100 to perform or use the exemplary methods described herein. In some alternative implementations, the machine-readable storage medium 120 may include mass storage, removable storage, volatile read-write memory, read-only memory (ROM), and the like, or any combination thereof. Exemplary mass storage devices may include magnetic disks, optical disks, solid state disks, and the like. Exemplary removable memory may include flash drives, floppy disks, optical disks, memory cards, compact disks, tape, and the like. Exemplary volatile read-write memory can include Random Access Memory (RAM). Exemplary RAM may include active random access memory (DRAM), double data rate synchronous active random access memory (DDR SDRAM), passive random access memory (SRAM), thyristor random access memory (T-RAM), zero capacitance random access memory (Z-RAM), and the like. Exemplary read-only memory may include mask read-only memory (MROM), programmable read-only memory (PROM), erasable programmable read-only memory (PEROM), electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM), digital versatile disk read-only memory, and the like. In some alternative implementations, the machine-readable storage medium 120 may be implemented on a cloud platform. For example only, the cloud platform may include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, etc., or any combination thereof.

In a specific implementation, the plurality of processors 110 execute computer executable instructions stored by the machine-readable storage medium 120, so that the processors 110 may execute the method for detecting multimedia data in combination with RPA according to the method embodiment, where the processors 110, the machine-readable storage medium 120, and the communication unit 140 are connected through the bus 130, and the processors 110 may be used to control the transceiving actions of the communication unit 140.

The specific implementation process of the processor 110 may refer to the above-mentioned method embodiments executed by the computer device 100, and the implementation principle and technical effects are similar, which are not described herein again.

In addition, the embodiment of the application also provides a readable storage medium, wherein computer executable instructions are preset in the readable storage medium, and when a processor executes the computer executable instructions, the method for detecting the multimedia data by combining with the RPA is realized.

Similarly, it should be noted that in order to simplify the description of the present disclosure and thereby aid in understanding one or more embodiments of the application, various features are sometimes grouped together in a single embodiment, figure, or description thereof. Similarly, it should be noted that in order to simplify the description of the present disclosure and thereby aid in understanding one or more embodiments of the application, various features are sometimes grouped together in a single embodiment, figure, or description thereof.

Claims

1. A method for detecting multimedia data in combination with RPA, the method being implemented by a computer device running an automated RPA system with a robotic process, the method comprising:

2. The method for detecting RPA-combined multimedia data according to claim 1, wherein the performing differential training on the second anomaly violation detection network according to the multimedia interactive image sequence carrying the prior anomaly violation feature data, the multimedia interactive image sequence not carrying the prior anomaly violation feature data, and the corresponding first fuzzy anomaly violation feature data, to generate a final target anomaly violation detection network comprises:

3. The method according to claim 2, wherein generating the first fuzzy anomaly violation feature data of the multimedia interaction image sequence that does not carry a priori anomaly violation feature data based on the second anomaly violation detection network and the multimedia interaction image sequence that does not carry a priori anomaly violation feature data comprises:

4. The method for RPA-combined multimedia data detection according to claim 3, wherein performing abnormal violation feature detection on the first unsupervised interaction image according to the second abnormal violation detection network, generating first fuzzy abnormal violation feature data and a corresponding confidence level of the first unsupervised interaction image, comprises:

wherein the method further comprises:

5. The RPA-combined multimedia data detection method according to claim 4, further comprising:

according to the unsupervised interactive images of the current training rounds and the corresponding second fuzzy anomaly violation feature data and the multimedia interactive image sequences carrying the priori anomaly violation feature data of the current training rounds, iterating the third anomaly violation detection network to carry out differential training until the fifth number of second supervised interactive images and the corresponding second fuzzy anomaly violation feature data are adjusted by the network more than once, generating a fourth anomaly violation detection network, and generating the final target anomaly violation detection network;

Performing abnormal violation segmentation data monitoring on the second unsupervised interaction image according to the third abnormal violation detection network, and generating second fuzzy abnormal violation feature data and corresponding confidence of the second unsupervised interaction image, wherein the method comprises the following steps of:

wherein the method further comprises:

and when the third anomaly detection network is iterated for differential training according to the unsupervised interaction image of the current training round and the corresponding second fuzzy anomaly violation feature data and the multimedia interaction image sequence carrying the priori anomaly violation feature data of the current training round, optimizing the first fully-connected output unit based on the third anomaly detection network after cyclic training until the previous sixth number of second supervised interaction images and the corresponding second fuzzy anomaly violation feature data are adjusted by the network more than once, generating a second fully-connected output unit.

6. The method according to claim 5, wherein the third anomaly detection network performs differential training according to the multimedia interactive image sequence carrying the prior anomaly violation feature data, the multimedia interactive image sequence not carrying the prior anomaly violation feature data, and the corresponding second fuzzy anomaly violation feature data, and generates the final target anomaly violation detection network, further comprising:

7. The method for RPA-combined multimedia data detection according to claim 3, wherein performing abnormal violation feature detection on the first unsupervised interaction image according to the second abnormal violation detection network, generating a confidence level of first blurred abnormal violation feature data of the first unsupervised interaction image, comprises:

8. The method for RPA-combined multimedia data detection according to claim 3, wherein performing abnormal violation feature detection on the first unsupervised interaction image according to the second abnormal violation detection network, generating a confidence level of first blurred abnormal violation feature data of the first unsupervised interaction image, comprises:

9. A multimedia data detection apparatus incorporating RPA, implemented by a computer device running an automated robot-flow RPA system, the apparatus comprising:

10. A computer device comprising a processor and a machine-readable storage medium having stored therein machine-executable instructions that are loaded and executed by the processor to implement the RPA-combined multimedia data detection method of any one of claims 1-8.